Experimental Settings

To validate the proposed algorithms based on experimental data, one large-scale test event took place on May 15th, 2017 at the TASS test facilities in Helmond, Netherlands. These tests were relying on an early version of the integrated physical proof of concept demon-strator developed in the HIGHTS project and involved a platoon consisting of 3 equipped cars driving in a row: TASS’ Prius car (as lead vehicle), Objective’s BMW (as 2nd vehicle)

Figure 7.5: Test vehicles involved in the first HIGHTS field trials carried out in Helmond:

Objective’s BMW, Tass’s Prius and Ibeo’s Passat (left to right).

5.56 5.57 5.58 5.59 5.6 5.61 5.62

longitude [deg]

51.455 51.46 51.465 51.47 51.475 51.48

latitude [deg]

Figure 7.6: Test site and vehicles’ trajectory in Helmond, Netherlands (original photo from Google Map).

and Ibeo’s Passat (as 3rd and last vehicle)(see Fig. 7.5). During these experiments, each vehicle was equipped with a singe-band GPS receiver, a RTK GPS receiver, an ITS-G5 platform (i.e., Cohda MK5) and a central Blackhole data logging PC, making two full rounds along the A270/N270 highway section. The followed route deliberately included a combination of straight and curvy sections for better representativity and for realistic assessment. The true positions of the vehicles were logged using a RTK GPS for reference purposes (ground truth). Figure 7.6 shows the test site and the followed trajectories.

Due to some problems in the GPS measurements collected at Objective’s vehicle during the trials, Ibeo’s vehicle has been selected as the “ego” vehicle under test (i.e., in charge of performing cooperative data fusion). The latter receives CAMs encapsulating RTK GPS data from both Objective and Tass’ vehicles, measures the corresponding RSSIs out of the received messages (IR-UWB devices were not yet integrated for V2V ranging in the demonstration platform by the time these first trials were conducted), and finally performs fusion with its own on-board GPS position to improve its position accuracy. Furthermore,

10¹ 10² 10³ distance [m]

-110 -100 -90 -80 -70 -60 -50

received power [dBm]

Figure 7.7: Pathloss measurements and approximate large-scale models. In the linear regression, n_p = 2.5 (path loss exponent) and σ_Sh = 3.7 dB (standard deviation of shad-owing).

it also tracks (i.e., updates) the neighboring RTK GPS information received in CAMs using mobility prediction since this information may be out-dated at the fusion time otherwise.

From a LDM perspective, this can also be viewed as an improvement in comparison with basic position awareness (in the sense the “ego” perception about its neighbors does not only rely on the CAMs but has been updated).

To calibrate the required large-scale path loss model, we have considered both the RSSI using Cohda MK5 and the distance between the two involved vehicles using their GPS RTK receivers. The result of the linear regression analysis is shown in Figure 7.7.

This path loss model will be used as the measurement model in the EKF-based fusion engine for CLoc. We use the EKF but not PF herein for some reasons. For this first field test followed by a real-time test later, we plan to implement the algorithm in a limited processing unit inside the Cohda MK5 but not in a connected PC in the vehicle as a starting point for the sake of simplicity. The Cohda MK5 has an integrated GPS inside so the fusion-based CLoc algorithm can access directly the GPS data as well as the RSSIs measured out of the received CAMs and the associated CAM data. On the other hand, the PF version will be implemented in the connected PC as soon as integrated process is optimized which is expected after this first test. Note that the fusion results based on EKF herein is generalized and comparable with PF.

7.3.2 Results

Figure 7.8(a) compares the performance of the CLoc method (i.e., fusing GPS and ITS-G5 RSSI) with that of both filtered and raw GPS positions. As it can be seen, the proposed CLoc approach outperforms the filtered GPS even though the localization accuracy gain is quite marginal and modest, as expected. This is likely due 1) to the very low number cooperative neighbors available in the test case (only 2, at most), 2) to very poor GDOP conditions, as the three vehicles were forming a “longitudinal” platoon most of the time and the “ego” vehicle considered for fusion was the leading one, and 3) to the relatively low CAM rate while providing RSSIs and neighboring positions, at approximately 3 Hz (in average) whereas a maximum 10 Hz could be used (i.e., nominal rate considered in most simulation-based evaluations of CLoc so far).

On the other hand, Figure 7.8(b) shows the performance associated with the LDM maintained at the IBEO’s “ego” vehicle (i.e., the quality and validity of the presumed neighbors’ positions). As expected, the prediction-based scheme achieves much higher localization accuracy than that without prediction. Specifically, the former performs pre-diction of neighboring vehicles based on their latest broadcast states (i.e., position and velocity) and a mobility model, whereas the latter simply relies on their raw positional information (i.e., communicated in the CAM). A closer look at this figure reveals that the accuracy gain is huge. Without prediction, the error accumulates quickly, especially when not receiving new CAMs due to too low CAM rate or simply packet loss. Moreover, higher position estimation rate (i.e., 8 Hz, as the GPS rate) would require an equivalent CAM rate to draw maximum benefits, which could not be met in these first experiments. Fig-ure 7.9 illustrates this observation, showing the RMSE of the position awareness regarding the 2 neighbors (Objective and Tass) over time. Note that the value on the right vertical axis CAM update takes either 0 if not receiving any CAM or 1 if receiving a CAM at any iteration. Overall, prediction globally improves position awareness about neighbors in the LDM by a factor of 10.

As aforementioned, the CAM rate of about 3 Hz is relative low when compared to the fusion rate of 8 Hz. Therefore, most of the iterations just correspond to filtered GPS but not to a true CLoc fusion event, leading to modest accuracy gains. To avoid this, we have performed other offline test, reducing the fusion rate down to 4 Hz, as shown in Figure 7.10.

The benefit of fusion-based CLoc over standalone GPS is thus more remarkable.

0 1 2 3 4 5

(b) LDM at the Ibeo’s “ego” vehicle.

Figure 7.8: Empirical CDFs of localization errors for the first trip of field trials in Helmond.

10 20 30

Figure 7.9: Localization RMSEs of the LDM at IBEO’s “ego” vehicle as a function of time for the first trip of field trials in Helmond. Cooperative awareness of Objective’s and Tass’ vehicles positions without prediction (top left and top right, respectively) versus with prediction (bottom left and bottom right, respectively).

The impact of GDOP on the CLoc accuracy has also been investigated. For this sake, the localization error vector has been projected onto the cross-track and along-track axes.

Considering the GDOP conditions in this test case (i.e., a platoon in line), the along-track errors are mostly improved by CLoc, as confirmed by Figure 7.11. The figure also shows that the cross-track errors are marginally improved. This is due to the fact that a

“longitudinal” platoon was maintained during most of the test.

During the tests in Helmond, the 3 vehicles drove for a second time on the same route (2nd trip). The results are summarized in Figure 7.12 and Figure 7.13. Interestingly, the CLoc method now improves quite significantly accuracy, especially in the lower error regime, as shown in Figure 7.12(a) and Figure 7.13. As the distances between the 3 vehicles

0 1 2 3 4 5 localization error [m]

0 0.2 0.4 0.6 0.8 1

empirical CDF(error)

raw GPS @ 4 Hz filtered GPS @ 4 Hz

fused GPS @ 4 Hz + ITS-G5 @ 3 Hz

Figure 7.10: Empirical CDFs of localization errors of the Ibeo’s “ego” vehicle for the first trip of field trials in Helmond with reduced position estimation rates.

0 0.5 1 1.5 2 2.5 3 3.5 4

along-track error [m]

0 0.5 1

empirical CDF(error)

0 0.5 1 1.5 2 2.5 3 3.5 4

cross-track error [m]

0 0.5 1

empirical CDF(error)

filtered GPS @ 4 Hz

fused GPS @ 4 Hz + ITS-G5 @ 3 Hz

Figure 7.11: Empirical CDFs of along-track and cross-track errors of the Ibeo’s “ego”

vehicle for the first trip of field trials in Helmond, with reduced position estimation rates.

were shorter during this second trip, RSSI measurements could contribute as more reliable and meaningful distance-dependent information to the final position estimates¹.

7.4 Summary

This chapter contributes to the validation of algorithms from our CLoc framework. On the one hand, relying on simulated mobility traces and assuming V2V IR-UWB range measurements, several observations can be made at the system level in view of the context-aware localization strategy.

1Theoretically, RSSI-based range measurements have standard deviation proportional to the true dis-tance.

0 1 2 3 4 5

(b) LDM at the Ibeo’s “ego” vehicle.

Figure 7.12: Empirical CDF of localization errors for the second trip of field trials in Helmond.

Figure 7.13: Empirical CDF of localization errors of the Ibeo’s “ego” vehicle for the second trip of field trials in Helmond with reduced position estimation rates.

• Fusion with other on-board sensors (i.e., WSS and IMU) is always beneficial, con-tributing mostly to control and stabilize the errors in the dimension along and or-thogonal to the road direction, regardless of environmental conditions;

• V2V cooperation is systematically beneficial, leading to sub-meter accuracy in WC error regimes and even 0.2 m accuracy in median error regimes, thus fulfilling the claimed applicative target;

• V2V cooperation is not necessarily useful if vehicle is equipped with a high-class GNSS by default (e.g., RTK and PPP), while operating in favorable conditions (i.e., open or intermediary urban environments);

• V2V cooperation rather strongly depends on the relative geometric configuration and connectivity conditions for isolated vehicles, for instance due to static NLOS situations (thus, leading to loose cooperative links) and/or due to “accordion” mo-bility pattern (e.g., when a peripheral node with respect to the rest of the VANET is stuck alone at an intersection red traffic light, whereas other vehicles ahead belong-ing to the same steady-state group have all turned already, thus leadbelong-ing to sparser connectivity and even poorer GDOP conditions). However, this shall be also mit-igated in real operating conditions. Especially, in dense urban environments (i.e., where the expected gain should be by the way larger in comparison with nominal GNSS), each vehicle possibly relies on a plurality of vehicles around itself (not even specifically belonging to a unique group moving in the same direction);

• Mobility-based prediction in CLoc, even when relying on simplistic model such as the bicycle model, looks fairly robust enough with respect to possible model mismatch in case of realistic urban mobility (e.g., with more erratic behavior than steady-state mobility regimes for instance on highways).

On the other hand, offline experimental validations in a highway scenario, while re-lying uniquely on GPS data and notoriously dispersed ITS-G5 V2V RSSI measurements as input observations, show already interesting gains through V2V cooperation beyond nominal GNSS/GPS performance. This is the case not only in terms of “ego” longitudinal localization, but also (and even more significantly) in terms of position awareness regard-ing neighborregard-ing vehicles through mobility-based predictions (i.e., enablregard-ing accurate LDM updates). It has been shown that the observed performance gains mostly depends on the rate of ITS-G5 messages broadcast (in average 3 Hz in the conducted tests, to be compared with 10 Hz for the “ego” onboard GPS rate), as well as on a relatively unfavorable GDOP (i.e., the three vehicles involved in the experiments being strictly aligned for the whole experiments). Furthermore, the V2I RSSI information available in the collected data set could not be fully exploitable, due to uncertain RSUs placement. Accordingly, higher V2X ITS-G5 transmission rates (up to 10 Hz), a better geo-referencing of static RSUs serving as anchors, a more realistic varying platoon topology over time, and finally the use of more accurate ranging-enabled technologies such as IR-UWB should be recommended in future field validations.

Conclusions and Perspectives

8.1 Conclusions

In this thesis, we have presented a Cooperative Localization (CLoc) framework for con-nected vehicles or vehicular ad hoc networks (VANETs), in which vehicles exploit the positioning capabilities of their neighbors and accordingly, enhance their own location es-timates. Due to its maturity (but also to its foreseen massive deployment in the short term), we have primarily chosen ITS-G5/IEEE 802.11p as main supporting vehicular com-munication technology¹. The general concept of CLoc, which has been covered rather ex-tensively in the literature in a variety of applications, may look promising in this vehicular context too at very first sight. However, as traditional CLoc techniques are adapted nei-ther to the VANET connectivity conditions nor to the experienced mobility patterns, their direct application is still non trivial and requires attention. Keeping these unprecedented challenges in mind, the main goal of this research work was to reach resilient sub-meter localization accuracy so as to meet the needs of Day-2 C-ITS applications. Our proposed solution has been tested through various sophisticated simulations and partly validated (offline) through experimental data from field tests. These validations have shown that the required level of accuracy could indeed be conditionally achieved (even in particularly pathological cases and in compliance with imposed standardization constraints), thanks to selective V2X cooperation and to multi-sensor fusion. The main contributions of this thesis can be summarized as follows.

1Note that our research methodology claims enough generality (e.g., aiming at the joint optimization of fusion algorithms and V2X transmission policy). Accordingly, it could get easily adapted to other relevant standards in turn (C-V2X such as LTE-V2X, 5G, etc.).

145

In Chapter 3, we have established a generic cooperative fusion framework based on a particle filter (PF) and adapted to the ITS-G5 communication technology. First, we have proposed prediction-based data resynchronization mechanisms to properly incorporate co-operative information incoming from asynchronous neighboring vehicles. This allows to mitigate possible biases in the neighboring position awareness, which must be injected into the fusion engine. We have also developed link selection mechanisms based on theoretical performance bounds so as to reduce complexity and minimize traffic (e.g., whenever cou-pled with a Tx censoring policy), without affecting significantly accuracy/latency. Results show for instance that the amount of required packets can be reduced by 70%, while loosing 14–18% of accuracy through selective fusion (in comparison with exhaustive fusion).

Chapter 4 adopts the same nominal framework as in Chapter 3 but it focuses more on the inherent specificities of V2V wireless connectivity (in terms of both propagation channel and communication channel congestion), evaluating and mitigating their impact-band (IR-UWB) ranging capabilities. On this occasion, we have shown that very poor initial GNSS prior information and/or unwanted error propagation induced by V2V coop-eration among vehicles could prevent from drawing maximum benefits from very accurate ranging, or could even lead to filter overconfidence in biased results and thus, to global divergence. Applying fusion scheduling and/or adaptive observation noise dithering to our CLoc algorithms, we have observed that when the biases are correctly mitigated (i.e., avoiding error propagation between vehicles and avoiding filter overconfidence in too poor estimates), the GNSS+IR-UWB fusion scheme then outperforms any other CLoc algo-rithm and naturally, also the standalone GNSS receiver option. On the one hand, under heterogeneous GNSS conditions/classes at the cooperating vehicles, fusion scheduling has been shown to provide an accuracy of 0.4 cm with 95% probability (compared to 25%

for conventional GNSS+IR-UWB fusion schemes). On the other hand, adaptive dither-ing achieves 0.2 m accuracy with 90% probability (compared to 48% for conventional GNSS+IR-UWB fusion schemes) in homogeneous GNSS capabilities.

In Chapter 6, we have proposed a hybrid V2X multisensor CLoc scheme, which requires additional on-board sensors (e.g., inertial or odometry sensors), camera-based lane detec-tor, etc. and even possibly, fixed elements of infrastructure (e.g., road side units (RSUs)).

The fusion with other on-board sensors (typically, WSS and IMU) has been shown always beneficial, contributing mostly to control and stabilize the error in the dimension

orthog-onal to the road direction. In tunnel scenarios, facing even more critical problems of fast divergence, we have proposed guidelines to apply hybrid CLoc with generalized V2X mea-surements. Considering more particularly V2X IR-UWB measurements (i.e., with respect to both mobile vehicles and RSUs), our CLoc solution can thus achieve median errors of 0.2 m approximately. The latter is also more attractive than CLoc assisted by GNSS repeaters in terms of both accuracy and cost of deployment. Finally, whenever ITS-G5 RSUs are used instead of IR-UWB enabled RSUs, we have shown they must be massively deployed (say, with less than 100 m as inter-side RSU interval) and thus, become costly.

In Chapter 7, results are first presented using a large-scale urban scenario that offers mixed environmental characteristics in view of GNSS performance (i.e., spanning from open environments to urban canyon), considering realistic mobility traces generated by a devoted traffic simulator (SUMO). We have also shown that, even in challenging se-tups (e.g., occasionally poor connectivity conditions and poor relative geometry), it is still possible to achieve 0.2 m accuracy with probability of 50%. One step ahead, we have performed offline validations using experimental data from a small-scale field test (3 vehicles only), relying uniquely on GPS data and notoriously dispersed IST-G5 V2V RSSI measurements as input observations. On this occasion, despite a quite restrictive scenario, we have already shown interesting gains through V2V cooperation, at least sig-nificantly beyond nominal GPS performance. This is the case not only in terms of “ego”

longitudinal localization, but also (and even more significantly, by about 10x) in terms of position awareness regarding neighboring vehicles through mobility-based predictions (i.e., enabling accurate LDM updates).

To summarize, this comparative study has shown that a sub-meter accuracy is overall possible through CLoc. We have also given practical guidelines for the design of future CLoc systems, thus contributing to the development of reliable and accurate location-based services for C-ITS.